Are we Bayesian referring expression generators

Abstract

A recent paper by Frank and Goodman (2012) proposes a Bayesian model of simple referential games. One of the claims embodied in the model is that choosing which word or property to use to refer to an object depends on the utility of the property. In this paper, we compare this model to other computational models of reference production, in particular the recent pro (Probabilistic Referential Overspecification) model. We argue that the assumption of utility that guides property choice in the Frank and Goodman (2012) model is inadequate, insofar as it ignores the possibility of overspecification and the role of preference rankings among properties, as a result of which they may be used irrespective of their utility. We show that models that do take this into account, such as pro, have a better fit to experimental data in which participants have the possibility of overspecifying.